• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

J4 ›› 2011, Vol. 33 ›› Issue (4): 98-101.

• 论文 • 上一篇    下一篇

SOFM模型在杂草图像识别中的应用

王焱,王磊明,孙雁鸣   

  1. (辽宁工程技术大学电控学院,辽宁 葫芦岛 125100)
  • 收稿日期:2010-05-10 修回日期:2010-08-03 出版日期:2011-04-25 发布日期:2011-04-25
  • 作者简介:王焱(1970),女,江苏无锡人,博士,副教授,研究方向为模式识别、测控技术与仪器。王磊明(1983),男,河北石家庄人,硕士,研究方向为模式识别。孙雁鸣(1984),男,吉林长春人,博士,研究方向为测控技术与仪器。

Application of the SelfOrganizing Feature Map (SOFM) Neural Network Model in Weed Identification

WANG Yan,WANG Leiming,SUN Yanming   

  1. (School of Electrical and  Control Engineering,Liaoning Technical University,Huludao 125100,China)
  • Received:2010-05-10 Revised:2010-08-03 Online:2011-04-25 Published:2011-04-25

摘要:

针对在杂草图像分割方面存在使用阈值分割需要选择分割阈值、图像分割精度不高等不足,本文结合超绿特征分割算法和SOFM网络,构造出一种杂草图像识别模型——GSOFM空间聚类模型。该方法是一种无监督学习方式,不需要指定阈值,利用网络自组织、自竞争的特性,实现对杂草图像的分割。在对图像进行超绿特征处理之后,使用超绿特征的灰度和归一化两个特征向量,实现SOFM空间聚类。实验结果表明,改进的GSOFM 方法相比其他三种杂草图像分割算法的分割结果都有一定的提高,分别比HIS阈值分割、超绿特征分割、双阈值分割提高28%、20%、21%。本算法结合后期形态学去噪后,识别正确率可达94%。

关键词: SOFM, 杂草识别, 图像分割, 超绿特征

Abstract:

As for the weeds image recognition model, the GSOFM spacial clustering model is developed especially for the shortage that segmentation threshold should be selected on the weeds image segmentation by threshold segmentation, which is combined with the segmentation algorithm of super green features and the SOFM network. This method is an unsupervised learning way without a specified threshold, which realizes the weeds image segmentation via the network’s characteristics of selforganization and competition. However, SOFM spatial clustering is achieved through the two eigenvectors of gray scale and normalization in the super green feature after processing.The experimental results show that the segmentation results have got some certain improvement with the improved GSOFM method, compared with other three kinds of weeds image segmentation algorithms, rising by 25%, 30% and 28% respectively than the HIS threshold segmentation, the super green characteristics segmentation and the double thresholds segmentation. The identification accuracy can reach 94% with this algorithm after denoising which is combined with morphology in the later phase.

Key words: SOFM;weed identification;image segmentation;extragreen feature